Enhancement request: particle selection using lasso tool in UMAP

[nkjaerjensen]

Hi,

I have noticed the cryoDRGN_filtering.ipynb jupyter notebooks in the output/analyze.epoch folders. In cryoDRGN, it is possible for the user to interactively select particle clusters from the UMAP using a lasso. I briefly attempted to use the notebook for the opusTOMO outputs, but it didn't seem to be compatible. Are there plans to make this lasso tool compatible with opusTOMO for focused training on a subset of particles? Or is it more in the spirit of opusTOMO to avoid this and do focused analysis on selected classes, outputting both conformation and composition latent codes?

[alncat]

Hi @nkjaerjensen , thank you very much for the feedback! It is a good idea to have a more graphical and interactive interface for selecting particles. I will take some time to test the notebook.

By now, If you want do focused training on a subset of particles, suppose your original starfile is named all.star, you can extract the starfile for the corresponding classes from the original one using the command,

dsd parse_pose_star all.star -D 128 --Apix 3.37 --labels analyze.29/kmeans20/labels.pkl --outdir analyze.23/kmeans20/

The distribution of classes in UMAP visualization is plotted in the figure analyze.29/kmeans20/umap.png and analyze.29/kmeans20/umap_hexbin.png.
The subtomograms for each class in that UMAP are stored under the outdir.
Remember to change the -D and --Apix according to the settings of your data.

You can combine some classes to form a bigger cluster using

dsdsh combine_star analyze.29/kmeans20/pre8.star analyze.29/kmeans20/pre9.star sel.star

which combines class8 and class9 into sel.star. Then, use

dsdsh sel.star analyze.29/kmeans20/pre10.star sel.star

to iteratively combine more classes.

You can then use sel.star to do a focused training on the selected subset~

Currently, the limitation is that we can only select particles using classses defined by KMean clustering.

[alncat]

By the way, if you want to reconstruct the conformation for a selected subset, except the conventional backprojection method, you may also try the fixed mode in OPUS-TOMO.

Details are given in https://github.com/alncat/opusTomo/wiki/Working-with-Warp .

By setting the encode-mode to be fixed, the subtomograms in that subset are assumed having the same latent code. By training opus-tomo, you can get a subtomogram average as conventional backprojection. But we are iteratively minimizing reconstruction loss instead of using simple backprojection here.

[alncat]

Hey @nkjaerjensen, thanks again for your great suggestions!

I have now tested and patched the cryoDRGN_filtering.ipynb under the analyze.N folder. The main selection functions excepct the lasso tool now work. You may need install jupyter notebook, plotly, and ipywidgets in the opustomo environment to work with the notebook. I installed them using the command pip install notebook plotly ipywidgets.

The update notebook is here https://github.com/alncat/opusTomo/blob/main/cryodrgn/templates/cryoDRGN_filtering_template.ipynb

Now, you can get more visualizations in cryodrgn style by running dsdsh analyze . N 10 20 first, then working with the cryoDRGN_filtering.ipynb. Remeber to remove the old cryoDRGN_filtering.ipynb in the analyze.N folder first before analyzing, then the new version of notebook will be copied to the analyze.N folder. Most of plotting and filtering operations now work as you can see in the notebook. In contrast to cryodrgn, I chose to save the selection as a starfile by default, which contains more metadata like coordinates in tomogram and can be visualized using other tools like artiax.

Lastly, I still need some time to test and patch the lasso tool, which depends on several IO related classes and will take me some time to make them cooperate.

[alncat]

@nkjaerjensen I also take some time to develope a set of skills for ai assisted analysis. You may check https://github.com/alncat/opusTomo?tab=readme-ov-file#skill . Currently, I am using it with https://moonshotai.github.io/kimi-cli/en/ . This set of skills covers some basic analyses. I will update them~ It seems like a feasible step towards automatic analyses.

[alncat]

@nkjaerjensen , the lasso selection tool now works~ you may need to install anywidget via pip install anywidget first. Next, i may add some analyses about the conformational latent space.

[nkjaerjensen]

Thanks so much for all the swift updates and new features @alncat! Super exciting to have these tools, I will let you know if I run into any bugs.

[alncat]

@nkjaerjensen you are welcome! just let me know if you run into any problems.

[alncat]

@nkjaerjensen I have added new capacities for visualizing a selected tomogram and selected particles in that tomogram to the notebook~